File size: 1,549 Bytes
a12256d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
---

license: cc-by-4.0
datasets:
- openslr/librispeech_asr
language:
- en
pipeline_tag: audio-to-audio
---


# SSLZip

## Usage

```py

import onnxruntime as ort

from transformers import HubertModel

import torch



# Load the upstream HuBERT model.

upstream = HubertModel.from_pretrained("facebook/hubert-base-ls960")

upstream.eval()



# Load the autoencoder model.

postprocessor = ort.InferenceSession("sslzip_256.onnx")

node_name = postprocessor.get_inputs()[0].name



# Prepare an input waveform (assuming 16kHz audio).

x = torch.randn(1, 16000)



# Extract the latent representation for downstream tasks.

with torch.inference_mode():

    h = upstream(x, output_hidden_states=True).hidden_states[-1]

    z = postprocessor.run(None, {node_name: h.cpu().numpy()})[0]



# Use z as you like.

print(z.shape)

```

## License

The pretrained model was developed using the LibriSpeech corpus and is distributed under the same license (CC BY 4.0).  
Please include credit to Nagoya Institue of Technology and Techno-Speech, Inc. when using this model.

## Citation

```bibtex

@InProceedings{yoshimura2025sslzip,

  author = {Takenori Yoshimura and Shinji Takaki and Kazuhiro Nakamura and Keiichiro Oura and Takato Fujimoto and Kei Hashimoto and Yoshihiko Nankaku and Keiichi Tokuda},

  title = {{SSLZip}: Simple autoencoding for enhancing self-supervised speech representations in speech generation},

  booktitle = {13th ISCA Speech Synthesis Workshop (SSW 2025)},

  pages = {xxx--xxx},

  year = {2025},

}

```